This is an early release preview. You may encounter bugs.
Collective Intelligence logo
Unclaimed

Tool inference

Collective Intelligence

Point an AI agent at your company database and it writes SQL that looks right and gives the wrong answer, because it doe

B 74/100 GitHub score ? This grade is derived from GitHub signals, not user votes. Open for the full breakdown.
No votes yet

01 / About

What Collective Intelligence is.

Ailin¹ Collective Intelligence is an orchestration engine that coordinates many models on one request instead of routing to a single one. It presents an OpenAI-compatible API, so a client changes only its base URL, and resolves each request to one of 32 registered strategies: a single cheapest-viable model when speed matters, consensus with a deterministic verifier when a machine-checkable answer is declared, or an explicitly named strategy such as blind debate, expert panels, devil's-advocate consensus, cost-cascade, or best-of-N.

A continuous discovery engine indexes 103,224 model identifiers across roughly 90 provider integrations rather than a hardcoded list, analysing each for capabilities, performance profile, pricing, context window, modalities, and architecture, and advertising a route only after it is proven live. Team assembly then elects models per request from that catalogue against the task, the strategy, and an outcome profile. Two independent request fields bound the result: maxcost sets a ceiling and qualitytarget sets a floor.

Every response carries ailin_metadata — the strategy used, the models called, the final decider, per-subcall cost, and dissent. The project publishes a benchmark on objectively graded tasks with the per-execution data and regeneration scripts committed alongside it:

System Objective accuracy
Ailin¹ collective, verified consensus 97% (37 of 38)
GPT-5.5-pro, Claude Opus 4.8, Gemini 3.1 Pro, Grok 4.3 68–82% pooled

The same write-up records where the approach does not currently lead: single models still win on open-ended prose such as creative writing and refactoring, the collective carries a cost premium except when the verifier short-circuit fires, and multi-round arbitration adds latency. Documented non-fits are high-volume low-stakes traffic, tight latency budgets, and documentation-style prose.

Features

  • OpenAI-compatible API: chat, responses, embeddings, images, and files, reached by swapping the base URL of any OpenAI SDK or curl client
  • 32 orchestration strategies: consensus with convergence floors, blind debate, expert panels, devil's-advocate consensus, cost-cascade, and best-of-N with objective verification, each labelled auto-selectable, explicit-only, or roadmap
  • ailin-auto cascade: five chat aliases span an instant single model to five-model consensus, with the cascade defaulting to the cheap path and escalating only when the quality gate demands it
  • Discovery engine: scans native provider APIs, cloud hubs, aggregators, open-model repositories, and private inference endpoints, absorbing new models with no code change or downtime
  • Deterministic verifier: arms when a request declares ailinconstraints.answercheck, and short-circuits the collective when the answer is machine-checkable
  • Decision provenance: ailin_metadata records strategy, models, decider, per-subcall cost, and dissent on every response
  • Cost and tenant governance: per-request max_cost enforced at admission, with architectural tenant isolation and quota controls at the platform layer
  • Multimodal and file output: image, audio, and video generation routed by capability, plus deterministic DOCX, XLSX, PDF, PPTX, ZIP, and code rendering from any structured-output chat model
  • Self-hosted stack: Docker Compose brings up the API, Postgres, and Redis, with a degraded mode that boots against a local Ollama and no provider key
  • Reproducible benchmark: raw per-execution CSVs and scripts are committed so every published table can be regenerated
  • License endpoints: /source and /license are served by the engine itself, with release provenance through SLSA, Sigstore, and an SPDX SBOM

02 / Discussion CREDIBILITY-GATED

Discussion

Reading is open to everyone. Posting and voting need a verified identity or a GitHub grade of B or higher.

  • No discussions yet.

03 / Build

Build with Collective Intelligence.

Browse the catalogue for frameworks, tools, and harnesses, each scored on real GitHub credibility.

Get Collective Intelligence →

Browse the catalogue